High storage and computational costs obstruct deep neural networks to be deployed on resource-constrained devices. Knowledge distillation (KD) aims to train a compact student network by transferring knowledge from a larger pretrained teacher model. However, most existing methods on KD ignore the valuable information among the training process associated with training results. In this article, we provide a new collaborative teaching KD (CTKD) strategy which employs two special teachers. Specifically, one teacher trained from scratch (i.e., scratch teacher) assists the student step by step using its temporary outputs. It forces the student to approach the optimal path toward the final logits with high accuracy. The other pretrained teacher (i.e., expert teacher) guides the student to focus on a critical region that is more useful for the task. The combination of the knowledge from two special teachers can significantly improve the performance of the student network in KD. The results of experiments on CIFAR-10, CIFAR-100, SVHN, Tiny ImageNet, and ImageNet datasets verify that the proposed KD method is efficient and achieves state-of-the-art performance.

Remote sensing image (RSI) scene classification aims to identify the semantic categories of remote sensing images based on their contents. Owing to the strong learning capability of deep convolutional neural networks (CNNs), RSI scene classification methods based on CNNs have drawn much attention and achieved remarkable performance. However, such outstanding deep neural networks are usually computationally expensive and time-consuming, making them impossible to apply on resource-constrained edge devices, such as the embedded systems used on drones. To tackle this problem, we introduce a novel pair-wise similarity knowledge distillation method, which could reduce the model complexity while maintaining satisfactory accuracy, to obtain a compact and efficient deep neural network for RSI scene classification. Different from the existing knowledge distillation methods, we design a novel distillation loss to transfer the valuable discriminative information, which could reduce the within-class variations and restrain the between-class similarity, from the cumbersome model to the compact model. This method could obtain the compact student model with higher performance compared with existing knowledge distillation methods in RSI scene classification. To be specific, we distill the probability outputs between sample pairs with the same label and match the probability outputs between the teacher and student models. Experiments on three public benchmark datasets for RSI scene classification, i.e., AID, UCMerced, and NWPU-RESISC datasets, verify that the proposed method could effectively distill the knowledge and result in a higher performance.

Different scales of the objects pose a great challenge for the segmentation of remote sensing images of special scenes. This paper focuses on the problem of large-scale variations of the target objects via a dynamical receptive field of the deep network. We construct a Gaussian dynamic convolution network by introducing a dynamic convolution layer to enhance remote sensing image understanding. Moreover, we propose a new Gaussian pyramid pooling (GPP) for multi-scale object segmentation. The proposed network can expand the size of the receptive field and improve its efficiency in aggregating contextual information. Experiments verify that our method outperforms the popular semantic segmentation methods on large remote sensing image datasets, including iSAID and LoveDA. Moreover, we conduct experiments to demonstrate that the Gaussian dynamic convolution works more effectively on remote sensing images than other convolutional layers.